摘要
浮选矿浆相气泡图像是采集自浮选槽内部矿浆溶液中的图像数据,与浮选泡沫相图像数据相比视觉特征显著不同。针对使用特殊设备从浮选槽矿浆溶液中原位采集的气泡图像数据,提出了一种基于DA-Attention U-Net编码-解码结构的气泡分割模型。模型以U-Net为基础,引入CBAM模块并依据Residual残差连接思想改进模块结构,使模型同时具有通道注意力和空间注意力的优点,给予包含气泡的前景区域更大权重,减少因下采样次数多导致的信息丢失;引入ASPP模块并基于Dense密集连接思想进行改进,从多尺度提取气泡特征及整合前后特征层信息;并在完成气泡分割的基础上使用热力图与显著图对分割结果进行分析。研究结果表明,与原始U-Net相比,所提模型对气泡图像分割效果更优,训练损失、Dice系数降低了0.416、0.2,分别达到了0.015、0.12,MIoU精度值、F1_Score值提升了0.331、0.229,分别达到了0.952、0.985,并通过消融试验验证了各模块有效性。该模型对气泡图像的精确分割,可为后续提取气泡特征奠定基础,对于未来将矿浆相气泡特征信息用于浮选过程智能控制,具有重要意义。
Flotation pulp phase bubble images are image data collected from the pulp inside the flotation cell and have significantly different visual characteristics compared to the flotation froth image data.A DA-Attention U-Net bubble segmentation model is proposed for bubble image data collected in situ from the flotation cell pulp solution using special equipment.This model improved the U-Net model by deploying the CBAM module with adjustments to the module structure based on the method of residual connection.As a result,the model owns both advantages in channel attention and spatial attention,assuring more weight is given to the foreground areas which contain bubbles in the image,and less information loss caused by the number of down-sampling.By deploying ASPP module with improvement based on the idea of Dense Connection,bubble features from multiple scales can be extracted,the information of front and back feature layers integrated,and the segmentation results analyzed using heat map and significance map.The research found that compared with the U-Net,the proposed model is more effective for bubble image segmentation,the training loss and Dice coefficient are reduced by 0.416and 0.2,reaching 0.015and 0.12respectively,and the MIoU accuracy value and F1_Score value are improved by 0.331and 0.229,reaching 0.952and 0.985respectively.Ablation experiments were conducted to verify the effectiveness of each module.The model′s accurate segmentation of bubble images can lay the foundation for subsequent extraction of bubble features,which is important for future research on the use of pulp phase bubbles′information for optimal control of flotation processes.
作者
徐宏祥
李神舟
徐培培
XU Hongxiang;LI Shenzhou;XU Peipei(School of Chemical&Environmental Engineering,China University of Mining and Technology Beijing,Beijing 100083,China;BGRIMM Machinery and Automation Technology Co.,Ltd.,Beijing 100160,China;BGRIMM Technology Group,Beijing 100160,China)
出处
《有色金属（选矿部分）》
CAS
2024年第6期106-115,131,共11页
Nonferrous Metals（Mineral Processing Section）
基金
国家重点研发计划项目(20212902702)。
关键词
浮选矿浆相气泡
语义分割
密集连接机制
注意力集中机制
浮选过程智能控制
flotation pulp phase bubbles
semantic segmentation
dense connection mechanism
attention focus mechanism
intelligent control of flotation process
作者简介
徐宏祥(1986-),男,山东东营人,博士,副教授,博士生导师,主要从事浮选理论及工艺研究。
